How is AI Control architected?
AI Control is a system of systems with two layers. The real-world control layer — ControlCore, identity, permission, policy, the AI Firewall, communication intelligence, the risk engine, gateways, edge nodes and independent safety — decides and enforces in real time. A mandatory Trust Chain layer on Solieum-Solana Layer 2 records verifiable proofs and never gates real-time safety.
- 13 components on the real-world control path.
- Trust Chain is mandatory (Solieum-Solana L2) and asynchronous.
- Every component fails to a safe state.
- Enforcement happens in gateways, never in the requesting AI.
Frequently asked questions
What is the two-layer architecture?
A real-time control layer that makes and enforces decisions, and a mandatory trust layer on Solieum-Solana Layer 2 that records proofs of identity, authorization, policy and revocation for independent verification.
Why is the Trust Chain kept out of the real-time path?
It is mandatory but asynchronous, so that trust-infrastructure latency or outages can never delay or weaken a safety decision. Local policy and physical safety continue if it is unavailable, and records anchor once it returns.
Where are decisions enforced?
At enforcement points — execution gateways, the AI Firewall, the Robot Gateway and edge nodes. The requesting AI is never trusted to enforce its own decision.